What agentic AI is

An agent is a program that decides its own next step. You hand it a goal instead of a script, and it works out how to get there — calling tools, reading what comes back, and trying again when something fails. That deciding, choosing the next move rather than following a fixed one, is what people mean by agentic.

The difference shows up in a simple contrast. A plain chatbot answers and stops; ask it to book a flight and it writes you the instructions. An agent, asked the same thing, checks the dates, searches, compares options, and comes back with a booking — or a clear reason it could not. The model underneath is the same kind of thing in both cases. What changes is that the agent runs in a loop and is allowed to act. An agent is not smarter than the model inside it; it is that model given tools and the permission to use them.

The loop

Almost every agent, however elaborate, is some version of four steps:

1. Observe — read the goal and whatever the world just returned.
2. Decide — pick the next action: which tool, with which arguments.
3. Act — call it: run code, query an API, search, write a file.
4. Repeat — feed the result back in and go again, until the goal is met or it gives up.

Tools are what make this useful. A model on its own can only produce text; give it a search tool, a way to run code, and access to files, and that same model starts doing real work. For example, a model that could only describe how to fix a bug becomes an agent that edits the file, runs the tests, reads the failure, and tries again. Memory — the notes it keeps between steps — is what lets it stay on a long task instead of forgetting what it already tried.

Where to go next

The rest of this handbook is the practical part: the patterns that keep agents on the rails, the tools and frameworks used to build them, and the step-by-step guides for assembling one — for example, a multi-machine stack that runs several agents at once. Start anywhere, or press / to search.